Non-Linear Book Manifolds: Learning from Associations the Dynamic Geometry of Digital Libraries

摘要

Mainstream approaches in the design of virtual libraries basically exploit the same ambient space as their physical twins. Our paper is an attempt to rather capture automatically the actual space on which the books live, and learn the virtual library as a non-linear book manifold. This tackles tantalizing questions, chief among which whether modeling should be static and book focused (e.g. using bag of words encoding) or dynamic and user focused (e.g. relying on what we define as a bag of readers encoding). Experiments on a real-world digital library display that the latter encoding is a serious challenger to the former. Our results also show that the geometric layers of the manifold learned bring sizeable advantages for retrieval and visualization purposes. For example, the topological layer of the manifold allows to craft Manifold association rules; experiments display that they bring dramatic improvements over conventional association rules built from the discrete topology of book sets. Improvements embrace each of the following major standpoints on association rule mining: computational, support, confidence, lift, and leverage standpoint.